humanizer

name	humanizer
version	1.1.0
description	Edit or review text to remove signs of AI-generated writing — inflated significance ('pivotal moment', 'evolving landscape'), promotional language ('nestled', 'breathtaking'), superficial -ing analyses ('highlighting', 'underscoring'), vague attributions ('industry experts'), em-dash overuse, rule-of-three patterns, AI vocabulary ('delve', 'tapestry'), copula avoidance ('serves as'), passive voice, negative parallelisms ('not just X, but Y'), filler phrases, sycophantic openings ('Great question!'), and the rest of the 29-pattern catalog from Wikipedia's 'Signs of AI writing'. Works on written prose only: essays, articles, drafts, documentation, emails, marketing copy. Use this skill whenever the user wants to humanize text, remove AI tells, edit AI-sounding writing, make text sound less like ChatGPT, do a final 'does this read like a person wrote it' pass, or hits you with phrases like 'humanize this', 'make this sound less AI', 'remove the AI slop', 'edit this draft', or 'why does this read like an LLM'.
license	MIT
compatibility	claude-code opencode
allowed-tools	["Read","Write","Edit","Grep","Glob"]
metadata	{"author":"MKAbuMattar"}

Humanizer

Edit text to remove the 29 most common AI-writing tells, then audit and revise until a human voice survives.

When to use

• The user pastes text and asks you to "humanize this", "make it sound less AI", "remove the AI slop", or "edit this draft".
• The user wants a final "does this read like a person wrote it" pass before publishing a post, README, email, or essay.
• The user wants you to review someone else's text and explain why it reads like an LLM.
• You yourself just generated a long-form piece (essay, blog post, announcement) and the user wants the polished version.

Skip this skill for: code comments, commit messages, log lines, structured data (JSON/YAML), tables, and short factual answers — they don't need a humanizing pass.

The loop in one paragraph

Read the source. Strip the 29 patterns (categories below). Read the result aloud in your head. Audit it by asking yourself "what makes the below so obviously AI generated?" — answer in 2-4 honest bullets. Revise to kill those remaining tells. Add soul (opinions, varied rhythm, specific details, "I" when it fits) so the result has a pulse instead of being merely clean. Output draft → audit → final.

Workflow

1. Read the source. If the user provided a writing sample for voice matching, read that first — note sentence length, word level, punctuation habits, recurring phrases. Load references/voice-calibration.md.
2. Identify patterns. Scan for all 29 tells. Load the reference file for each category you hit:
   • Patterns 1-6 (significance inflation, promotional language, -ing analyses, vague attributions, formulaic "Challenges" sections) → references/content-patterns.md
   • Patterns 7-13 (AI vocabulary, copula avoidance, negative parallelisms, rule of three, false ranges, passive voice) → references/language-patterns.md
   • Patterns 14-19 (em-dash overuse, boldface, inline-header lists, title case, emojis, curly quotes) → references/style-patterns.md
   • Patterns 20-22 (chatbot artifacts, knowledge-cutoff disclaimers, sycophantic tone) → references/communication-patterns.md
   • Patterns 23-29 (filler phrases, hedging, generic conclusions, hyphen pairs, authority tropes, signposting, fragmented headers) → references/filler-and-persuasion.md
3. Draft a rewrite. Replace each tell with a natural alternative. Keep the meaning; cut the puffery. Default to is/are/has over "serves as / functions as / stands as".
4. Add soul. A clean rewrite without voice still reads like an LLM. Load and inject opinions, rhythm variation, specific details, and a first-person take where it fits.